Automated Text Classification Using a Multi-Agent Framework

advertisement
Fu, Y., Ke, W., Mostafa, J. (2005). Automated Text Classification Using a Multi-Agent Framework. Proceedings of the 5th ACM/
IEEE-CS joint conference on Digital Libraries: Tools & Techniques track: supporting classification., pp. 157-158
Automated Text Classification Using a Multi-Agent
Framework
Yueyu Fu, Weimao Ke and Javed Mostafa
Laboratory of Applied Informatics Research
Indiana University, Bloomington
IN, 47405-3907
(812)856-4182, 01
{yufu, wke, jm}@indiana.edu
Also, a centralized classification system may be overwhelmed by a
large and dynamic document stream (e.g., online news).
ABSTRACT
Automatic text classification is an important operational problem in
digital library practice. Most text classification efforts so far
concentrated on developing centralized solutions. However,
centralized classification approaches often are limited due to
constraints on knowledge and computing resources. In addition,
centralized approaches are more vulnerable to attacks or system
failures and less robust in dealing with them. We present a decentralized approach and system implementation (named MACCI)
for text classification using a multi-agent framework. Experiments
are conducted to compare our multi-agent approach with a
centralized approach. The results show multi-agent classification
can achieve promising classification results while maintaining its
other advantages.
The Internet is a distributed system and it offers the opportunity to
take advantage of distributed computing paradigms and distributed
knowledge resources for classification. With this motivation we
attempted to develop a distributed automated classification
environment that can offer satisfactory classification performance.
Below, we describe the details of the system and experimental
studies conducted to evaluate the system on a standard document
collection, RCV1-v2 [2].
2.
MULTI-AGENT CLASSIFICATION
An agent is an autonomous computer program which can emulate
certain intelligent behavior and conduct tasks on behalf of its user.
Classification agents are independent homogeneous text classifiers
except that each agent can only classify content from a limited
domain. Agent coordination strategy, which controls agent
communication and interaction, is a key component of a multi-agent
system. When an agent is unable to identify any or all of the classes
for a document, the agent may seek help from other agents based on
the agent coordination strategy [4].
Categories and Subject Descriptors
H.3.4 [Information Storage and Retrieval]: Systems and Software
– distributed systems, performance evaluation
General Terms
Algorithms, Design, Experimentation
Agent representation: Documents and classes are represented in
Vector Space Model using TF*IDF term weights. Each agent is
represented as a feature vector. For each class, the features are the
top ranked terms from the corresponding training documents based
on TF*IDF term weights. Each document is represented as a
document vector. These terms are the top ranked terms from the
whole training set based on the TF*IDF term weights. Cosine
similarity score is calculated between a document vector and a class
vector. If the similarity score exceeds a pre-defined threshold, the
document is considered as a member of this class. Otherwise, this
classification fails.
Keywords
Classification, Multi-Agent System
1. INTRODUCTION
Text classification, broadly defined as determining and assigning
topical labels to content, is a fundamental operation in digital
libraries. Typically, automatic text classification has been conducted
in a centralized architecture [1, 2, & 3]. A single classifier is
responsible for classifying all the incoming documents.
Theoretically, a centralized classification system, which has all the
necessary knowledge and computing resources, can be built to solve
any classification problem. However, performing long-term
classification may exceed the capabilities of centralized
classification systems in practice. For example, the knowledge of a
centralized classification system may become “stale” over time.
Architecture: Multi-Agent Collaboration and Classification of
Information (MACCI), a multi-agent classification system has been
implemented using the DIET Agents platform [5]. There are two
kinds of agents in MACCI environment: an administration agent
which is in charge of distributing the documents from the document
pool and a group of classification agents which is responsible for
conducting the actual classification tasks.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
JCDL’05, June 7–11, 2005, Denver, Colorado, USA
Copyright 2005 ACM 1-58113-876-8/05/0006…$5.00.
2.1 Agent Coordination Strategy
A coordination strategy, called multi-agent Good-Neighbor strategy,
was developed for MACCI. The hierarchical structure of the classes
in the document set RCV1-v2 [2] was utilized to build the GoodNeighbor strategy. There are four parent classes at the top level.
Each of these classes has certain number of child classes. Some
157
(small) are presented below. Lewis and his colleagues conducted
experiments on the same document collection using different
centralized classification approaches [2]. Their benchmark result
using a SVM classifier was 0.82 (micro-averaging F1 ) and 0.61
(macro-averaging F1 ).
child classes have their own child classes. Each agent has two lists
of agents, called good neighbors, which can offer classification
service: success list and failure list. The success list contains the
agents that represent its parent and child classes. The failure list
contains the agents that represent the top level parent classes. Below
is the coordination algorithm:
1.
2.
3.
Table 1. Classification effectiveness
The administration agent distributes a document from the
document pool to a randomly chosen classification agent. This
step is repeated after certain interval until the document pool
is empty.
Method
If an agent successfully classifies a document, it sends the
document to the agents in its success list for other potential
classification. The help degree is set to 1.
5.
If an agent fails to classify a document, it sends the document
to agents in its failure list for help. The help degree is set to 1.
4.
If an agent successfully classifies a document sent from
another classification agent and the help degree is smaller than
4, it sends the document only to the agents that represent its
child classes in its success list. The help degree is incremented
by 1.
5.
If an agent fails to classify a document sent from another
classification agent, it doesn’t take any action. If none of the
agents can classify the document, this case is considered as a
NULL classification.
3.
EXPERIMENT DESIGN
Multi-agent
(small)
Multi-agent
(large)
microF1.0
0.72
0.57
0.51
macroF1.0
0.54
0.46
0.40
DISCUSSION & CONCLUSION
The micro-averaging F score is always higher than the macroaveraging F score in this case. According to Yang and Liu [1], the
micro-averaging score is more influenced by classification
performance on common classes and the macro-averaging score is
influenced by classification performance on rare classes. Since in the
RCV1-v2 training set the common classes have more positive
examples than the rare classes, our classifier performed better on the
common classes which resulted in higher micro-averaging F score.
Overall, the results show the multi-agent approach can provide
acceptable classification performance in terms of effectiveness,
which can be an alternative when a centralized approach is
impossible to realize. In the future, we will experiment with
additional coordination strategies such as a market-based strategy.
Efficiency should also be measured and further study about how to
balance between effectiveness and efficiency is needed.
A set of experiments has been conducted to compare the
performance of our multi-agent classification approach with a
centralized classification approach. A centralized classification
system was developed using the same classification algorithms as
the multi-agent classification approach.
6. ACKNOWLEDGMENTS
This work was partially supported through a grant from the National
Science Foundation Award#:0333623.
3.1 Data set
7. REFERENCES
RCV1-v2 [2] was chosen to be the data set for the experiments. It
contains a corpus of more than 800,000 manually categorized
newswire stories from Reuters, Ltd. The collection was split into a
training set of 23,149 documents and a test set of 781,265
documents. It contains documents from 103 classes. Some examples
of classes are corporation, economics, government, and markets. On
average, each document is a member of three classes.
[1] Yang, Y., and Liu, X. A re-examination of text categorization
methods. In Proceedings of the 22nd Annual International
ACM SIGIR Conference on Research and Development in
Information Retrieval, 42-49, 1999.
[2] Lewis, D. D.; Yang, Y.; Rose, T.; and Li, F. RCV1: A New
Benchmark Collection for Text Categorization Research.
Journal of Machine Learning Research, 5:361-397, 2004.
3.2 Evaluation Methodology
[3] Lewis, D. D. Evaluating and Optimizing Autonomous Text
Classification Systems. In Proceedings of the 18th Annual
International ACM SIGIR Conference on Research and
Development in Information Retrieval, 246-254, 1995.
The classification performance was evaluated using standard
effective measures including precision, recall, and F measure.
Micro-averaging and macro-averaging methods were used to
compute the average F scores. F1 is defined as follows:
F1 =
Centralized
[4] Mukhopadhyay, S., Peng, S., Raje, R., Palakal, M., & Mostafa,
J. Large-scale Multi-agent Information Classification Using
Dynamic Acquaintance Lists. Journal of the American Society
for Information Science & Technology, 54(10), 2003.
2 ∗ Pr ecision ∗ Re call
Pr ecision + Re call
[5] DIET. DIET Agent Platform, 2004.
http://diet-agents.sourceforge.net.
4. RESULTS
The experimental results from our centralized and multi-agent
approach were collected for comparison. The multi-agent
experiment on the entire test set (large) and a small proportion of it
158
Download